Skip to content

zhr1201/CNN-for-single-channel-speech-enhancement

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CNN-for-single-channel-speech-enhancement

A tensorflow implementation of the paper: A Fully Convolutional Neural Network for Speech Enhancement https://arxiv.org/abs/1609.07132 A processed sample can be found in audiosample/

Requirements

  • tensorflow r0.11
  • librosa
  • numpy

File documentation

  • SENN.py: The structure of the network.
  • audio_reader.py: Find the speech and noise in the files and enqueue the audios that have been read into tf.queue.
  • SENN_train.py: Train the net.
  • SENN_audio_eval: Use a noisy sample to evaluate the net/

Training procedure

  • Orgnize your clean speech files and noise files in different directories.
  • Change their dir in SENN_train.py and train the net.
  • Mix your own samples and test use SENN_audio_eval.py

Some other things

The original paper use per sample pre-whitening and we also use that in this piece of code, but it turns out better to use global mean and var to do the pre-whitening. We didn't use the skip connections and our tests show that the most important factor leading to good performance is the size of the noise data set. The model is very likely to overfit if only 100 types of noise are provided.

About

Convolutional neural nets for single channel speech enhancement

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages